[PERF] Large diff between HLL and nqp performance in code example

[zoffixznet]

Noticed a large difference between nqp version (3s) and HLL version (42s) of this code. Filing in case there are potential optimization opportunities waiting to be discovered.

With this input saved as day_05.input file

This runs in 42s and if you change all the .AT-POS to [], it goes up to 1m3s

my int @instructions = 'day_05.input'.IO.slurp.words.map: *.Int;
my int $index = 0;
my int $stop = @instructions.elems;
my int $iterations = 0;
my int $jump;

while $index < $stop {
    $jump = @instructions.AT-POS: $index;

    $jump >= 3 ?? --@instructions.AT-POS: $index
               !! ++@instructions.AT-POS: $index;

    $index = $index + $jump;
    ++$iterations;
}

say $iterations;

While this runs in 3s:

use nqp;

nqp::stmts(
  (my $ins := nqp::getattr('day_05.input'.IO.slurp.words.map(*.Int).eager.List,
    List, '$!reified')),
  (my int $index),
  (my int $stop = nqp::elems($ins)),
  (my int $jump),
  (my int $iterations),
  nqp::while(
    nqp::islt_i($index, $stop),
    nqp::stmts(
      ($jump = nqp::atpos($ins,$index)),
      nqp::if(
        nqp::isge_i($jump,3),
        nqp::bindpos($ins,$index,nqp::sub_i(nqp::atpos($ins,$index),1)),
        nqp::bindpos($ins,$index,nqp::add_i(nqp::atpos($ins,$index),1))),
      ($index = nqp::add_i($index,$jump)),
      ($iterations = nqp::add_i($iterations, 1)))),
  nqp::say($iterations));

[zoffixznet]

Don't see anything odd about it in --target=optimize or the profile. The biggest culprit is AT-POS, which is JITed (but not OSRed, which is already mentioned on #1421)